test_modeling_utils.py 31.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# coding=utf-8
# Copyright 2022 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Patrick von Platen's avatar
Patrick von Platen committed
16

patil-suraj's avatar
patil-suraj committed
17
import inspect
18
19
20
import tempfile
import unittest

21
import numpy as np
22
23
import torch

Patrick von Platen's avatar
Patrick von Platen committed
24
from diffusers import (
Patrick von Platen's avatar
Patrick von Platen committed
25
26
    BDDMPipeline,
    DDIMPipeline,
27
    DDIMScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
28
    DDPMPipeline,
29
    DDPMScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
30
    GlidePipeline,
Patrick von Platen's avatar
Patrick von Platen committed
31
32
    GlideSuperResUNetModel,
    GlideTextToImageUNetModel,
Patrick von Platen's avatar
Patrick von Platen committed
33
    GradTTSPipeline,
34
    GradTTSScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
35
    LatentDiffusionPipeline,
Patrick von Platen's avatar
Patrick von Platen committed
36
    NCSNpp,
Patrick von Platen's avatar
Patrick von Platen committed
37
    PNDMPipeline,
38
    PNDMScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
39
40
    ScoreSdeVePipeline,
    ScoreSdeVeScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
41
42
    ScoreSdeVpPipeline,
    ScoreSdeVpScheduler,
patil-suraj's avatar
patil-suraj committed
43
    UNetGradTTSModel,
anton-l's avatar
anton-l committed
44
45
    UNetLDMModel,
    UNetModel,
46
)
47
from diffusers.configuration_utils import ConfigMixin
Patrick von Platen's avatar
Patrick von Platen committed
48
from diffusers.pipeline_utils import DiffusionPipeline
49
from diffusers.pipelines.pipeline_bddm import DiffWave
Patrick von Platen's avatar
Patrick von Platen committed
50
from diffusers.testing_utils import floats_tensor, slow, torch_device
51
52


Patrick von Platen's avatar
Patrick von Platen committed
53
torch.backends.cuda.matmul.allow_tf32 = False
54
55


56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
class ConfigTester(unittest.TestCase):
    def test_load_not_from_mixin(self):
        with self.assertRaises(ValueError):
            ConfigMixin.from_config("dummy_path")

    def test_save_load(self):
        class SampleObject(ConfigMixin):
            config_name = "config.json"

            def __init__(
                self,
                a=2,
                b=5,
                c=(2, 5),
                d="for diffusion",
                e=[1, 3],
            ):
73
                self.register_to_config(a=a, b=b, c=c, d=d, e=e)
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88

        obj = SampleObject()
        config = obj.config

        assert config["a"] == 2
        assert config["b"] == 5
        assert config["c"] == (2, 5)
        assert config["d"] == "for diffusion"
        assert config["e"] == [1, 3]

        with tempfile.TemporaryDirectory() as tmpdirname:
            obj.save_config(tmpdirname)
            new_obj = SampleObject.from_config(tmpdirname)
            new_config = new_obj.config

Patrick von Platen's avatar
Patrick von Platen committed
89
90
91
92
        # unfreeze configs
        config = dict(config)
        new_config = dict(new_config)

93
94
95
96
97
        assert config.pop("c") == (2, 5)  # instantiated as tuple
        assert new_config.pop("c") == [2, 5]  # saved & loaded as list because of json
        assert config == new_config


patil-suraj's avatar
patil-suraj committed
98
class ModelTesterMixin:
99
    def test_from_pretrained_save_pretrained(self):
patil-suraj's avatar
patil-suraj committed
100
101
102
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
Patrick von Platen's avatar
Patrick von Platen committed
103
        model.to(torch_device)
patil-suraj's avatar
patil-suraj committed
104
        model.eval()
105
106
107

        with tempfile.TemporaryDirectory() as tmpdirname:
            model.save_pretrained(tmpdirname)
patil-suraj's avatar
patil-suraj committed
108
            new_model = self.model_class.from_pretrained(tmpdirname)
Patrick von Platen's avatar
Patrick von Platen committed
109
            new_model.to(torch_device)
110

patil-suraj's avatar
patil-suraj committed
111
112
113
        with torch.no_grad():
            image = model(**inputs_dict)
            new_image = new_model(**inputs_dict)
114

patil-suraj's avatar
patil-suraj committed
115
        max_diff = (image - new_image).abs().sum().item()
Patrick von Platen's avatar
Patrick von Platen committed
116
        self.assertLessEqual(max_diff, 5e-5, "Models give different forward passes")
117

patil-suraj's avatar
patil-suraj committed
118
    def test_determinism(self):
patil-suraj's avatar
patil-suraj committed
119
120
121
122
123
124
125
126
127
128
129
130
131
132
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()
        with torch.no_grad():
            first = model(**inputs_dict)
            second = model(**inputs_dict)

        out_1 = first.cpu().numpy()
        out_2 = second.cpu().numpy()
        out_1 = out_1[~np.isnan(out_1)]
        out_2 = out_2[~np.isnan(out_2)]
        max_diff = np.amax(np.abs(out_1 - out_2))
        self.assertLessEqual(max_diff, 1e-5)
133

patil-suraj's avatar
patil-suraj committed
134
    def test_output(self):
patil-suraj's avatar
patil-suraj committed
135
136
137
138
139
140
141
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output = model(**inputs_dict)
142

patil-suraj's avatar
patil-suraj committed
143
144
145
        self.assertIsNotNone(output)
        expected_shape = inputs_dict["x"].shape
        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
146

patil-suraj's avatar
patil-suraj committed
147
    def test_forward_signature(self):
patil-suraj's avatar
patil-suraj committed
148
149
150
151
152
153
154
155
156
        init_dict, _ = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        signature = inspect.signature(model.forward)
        # signature.parameters is an OrderedDict => so arg_names order is deterministic
        arg_names = [*signature.parameters.keys()]

        expected_arg_names = ["x", "timesteps"]
        self.assertListEqual(arg_names[:2], expected_arg_names)
157

patil-suraj's avatar
patil-suraj committed
158
    def test_model_from_config(self):
patil-suraj's avatar
patil-suraj committed
159
160
161
162
163
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()
164

patil-suraj's avatar
patil-suraj committed
165
166
167
168
169
170
171
        # test if the model can be loaded from the config
        # and has all the expected shape
        with tempfile.TemporaryDirectory() as tmpdirname:
            model.save_config(tmpdirname)
            new_model = self.model_class.from_config(tmpdirname)
            new_model.to(torch_device)
            new_model.eval()
172

patil-suraj's avatar
patil-suraj committed
173
174
175
176
177
        # check if all paramters shape are the same
        for param_name in model.state_dict().keys():
            param_1 = model.state_dict()[param_name]
            param_2 = new_model.state_dict()[param_name]
            self.assertEqual(param_1.shape, param_2.shape)
178

patil-suraj's avatar
patil-suraj committed
179
180
181
        with torch.no_grad():
            output_1 = model(**inputs_dict)
            output_2 = new_model(**inputs_dict)
182

patil-suraj's avatar
patil-suraj committed
183
        self.assertEqual(output_1.shape, output_2.shape)
patil-suraj's avatar
patil-suraj committed
184
185

    def test_training(self):
patil-suraj's avatar
patil-suraj committed
186
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
187

patil-suraj's avatar
patil-suraj committed
188
189
190
191
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.train()
        output = model(**inputs_dict)
192
        noise = torch.randn((inputs_dict["x"].shape[0],) + self.get_output_shape).to(torch_device)
patil-suraj's avatar
patil-suraj committed
193
194
        loss = torch.nn.functional.mse_loss(output, noise)
        loss.backward()
195

patil-suraj's avatar
patil-suraj committed
196
197
198
199
200
201
202
203
204
205
206
207
208

class UnetModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = UNetModel

    @property
    def dummy_input(self):
        batch_size = 4
        num_channels = 3
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor([10]).to(torch_device)

patil-suraj's avatar
patil-suraj committed
209
        return {"x": noise, "timesteps": time_step}
210

patil-suraj's avatar
patil-suraj committed
211
212
213
    @property
    def get_input_shape(self):
        return (3, 32, 32)
214

patil-suraj's avatar
patil-suraj committed
215
216
217
    @property
    def get_output_shape(self):
        return (3, 32, 32)
patil-suraj's avatar
patil-suraj committed
218
219
220
221
222
223
224
225
226
227
228

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "ch": 32,
            "ch_mult": (1, 2),
            "num_res_blocks": 2,
            "attn_resolutions": (16,),
            "resolution": 32,
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict
229

patil-suraj's avatar
patil-suraj committed
230
    def test_from_pretrained_hub(self):
patil-suraj's avatar
patil-suraj committed
231
232
233
        model, loading_info = UNetModel.from_pretrained("fusing/ddpm_dummy", output_loading_info=True)
        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)
patil-suraj's avatar
patil-suraj committed
234

patil-suraj's avatar
patil-suraj committed
235
        model.to(torch_device)
patil-suraj's avatar
patil-suraj committed
236
237
238
        image = model(**self.dummy_input)

        assert image is not None, "Make sure output is not None"
239

patil-suraj's avatar
patil-suraj committed
240
241
242
243
244
245
246
    def test_output_pretrained(self):
        model = UNetModel.from_pretrained("fusing/ddpm_dummy")
        model.eval()

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)
247

patil-suraj's avatar
patil-suraj committed
248
249
        noise = torch.randn(1, model.config.in_channels, model.config.resolution, model.config.resolution)
        time_step = torch.tensor([10])
250

patil-suraj's avatar
patil-suraj committed
251
252
        with torch.no_grad():
            output = model(noise, time_step)
253

patil-suraj's avatar
patil-suraj committed
254
255
        output_slice = output[0, -1, -3:, -3:].flatten()
        # fmt: off
Patrick von Platen's avatar
Patrick von Platen committed
256
        expected_output_slice = torch.tensor([0.2891, -0.1899, 0.2595, -0.6214, 0.0968, -0.2622, 0.4688, 0.1311, 0.0053])
patil-suraj's avatar
patil-suraj committed
257
258
259
        # fmt: on
        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))

260

Patrick von Platen's avatar
Patrick von Platen committed
261
262
class GlideSuperResUNetTests(ModelTesterMixin, unittest.TestCase):
    model_class = GlideSuperResUNetModel
patil-suraj's avatar
patil-suraj committed
263
264
265
266
267
268
269
270
271
272
273
274
275

    @property
    def dummy_input(self):
        batch_size = 4
        num_channels = 6
        sizes = (32, 32)
        low_res_size = (4, 4)

        noise = torch.randn((batch_size, num_channels // 2) + sizes).to(torch_device)
        low_res = torch.randn((batch_size, 3) + low_res_size).to(torch_device)
        time_step = torch.tensor([10] * noise.shape[0], device=torch_device)

        return {"x": noise, "timesteps": time_step, "low_res": low_res}
276

patil-suraj's avatar
patil-suraj committed
277
278
279
    @property
    def get_input_shape(self):
        return (3, 32, 32)
280

patil-suraj's avatar
patil-suraj committed
281
282
283
    @property
    def get_output_shape(self):
        return (6, 32, 32)
284

patil-suraj's avatar
patil-suraj committed
285
286
287
    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "attention_resolutions": (2,),
288
            "channel_mult": (1, 2),
patil-suraj's avatar
patil-suraj committed
289
290
291
292
293
294
295
296
            "in_channels": 6,
            "out_channels": 6,
            "model_channels": 32,
            "num_head_channels": 8,
            "num_heads_upsample": 1,
            "num_res_blocks": 2,
            "resblock_updown": True,
            "resolution": 32,
297
            "use_scale_shift_norm": True,
patil-suraj's avatar
patil-suraj committed
298
299
300
301
302
303
304
305
306
307
308
309
310
311
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict

    def test_output(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output = model(**inputs_dict)

        output, _ = torch.split(output, 3, dim=1)
312

patil-suraj's avatar
patil-suraj committed
313
314
315
        self.assertIsNotNone(output)
        expected_shape = inputs_dict["x"].shape
        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
316

patil-suraj's avatar
patil-suraj committed
317
    def test_from_pretrained_hub(self):
Patrick von Platen's avatar
Patrick von Platen committed
318
        model, loading_info = GlideSuperResUNetModel.from_pretrained(
319
320
            "fusing/glide-super-res-dummy", output_loading_info=True
        )
patil-suraj's avatar
patil-suraj committed
321
322
323
324
325
326
327
        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)

        model.to(torch_device)
        image = model(**self.dummy_input)

        assert image is not None, "Make sure output is not None"
328

patil-suraj's avatar
patil-suraj committed
329
    def test_output_pretrained(self):
Patrick von Platen's avatar
Patrick von Platen committed
330
        model = GlideSuperResUNetModel.from_pretrained("fusing/glide-super-res-dummy")
patil-suraj's avatar
patil-suraj committed
331
332
333
334

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)
335

336
        noise = torch.randn(1, 3, 64, 64)
patil-suraj's avatar
patil-suraj committed
337
338
        low_res = torch.randn(1, 3, 4, 4)
        time_step = torch.tensor([42] * noise.shape[0])
339

patil-suraj's avatar
patil-suraj committed
340
341
        with torch.no_grad():
            output = model(noise, time_step, low_res)
342

patil-suraj's avatar
patil-suraj committed
343
344
345
        output, _ = torch.split(output, 3, dim=1)
        output_slice = output[0, -1, -3:, -3:].flatten()
        # fmt: off
346
        expected_output_slice = torch.tensor([-22.8782, -23.2652, -15.3966, -22.8034, -23.3159, -15.5640, -15.3970, -15.4614, - 10.4370])
patil-suraj's avatar
patil-suraj committed
347
348
        # fmt: on
        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))
patil-suraj's avatar
patil-suraj committed
349

anton-l's avatar
anton-l committed
350

Patrick von Platen's avatar
Patrick von Platen committed
351
352
class GlideTextToImageUNetModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = GlideTextToImageUNetModel
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409

    @property
    def dummy_input(self):
        batch_size = 4
        num_channels = 3
        sizes = (32, 32)
        transformer_dim = 32
        seq_len = 16

        noise = torch.randn((batch_size, num_channels) + sizes).to(torch_device)
        emb = torch.randn((batch_size, seq_len, transformer_dim)).to(torch_device)
        time_step = torch.tensor([10] * noise.shape[0], device=torch_device)

        return {"x": noise, "timesteps": time_step, "transformer_out": emb}

    @property
    def get_input_shape(self):
        return (3, 32, 32)

    @property
    def get_output_shape(self):
        return (6, 32, 32)

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "attention_resolutions": (2,),
            "channel_mult": (1, 2),
            "in_channels": 3,
            "out_channels": 6,
            "model_channels": 32,
            "num_head_channels": 8,
            "num_heads_upsample": 1,
            "num_res_blocks": 2,
            "resblock_updown": True,
            "resolution": 32,
            "use_scale_shift_norm": True,
            "transformer_dim": 32,
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict

    def test_output(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output = model(**inputs_dict)

        output, _ = torch.split(output, 3, dim=1)

        self.assertIsNotNone(output)
        expected_shape = inputs_dict["x"].shape
        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")

    def test_from_pretrained_hub(self):
Patrick von Platen's avatar
Patrick von Platen committed
410
        model, loading_info = GlideTextToImageUNetModel.from_pretrained(
411
412
413
414
415
416
417
418
419
420
421
            "fusing/unet-glide-text2im-dummy", output_loading_info=True
        )
        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)

        model.to(torch_device)
        image = model(**self.dummy_input)

        assert image is not None, "Make sure output is not None"

    def test_output_pretrained(self):
Patrick von Platen's avatar
Patrick von Platen committed
422
        model = GlideTextToImageUNetModel.from_pretrained("fusing/unet-glide-text2im-dummy")
423
424
425
426
427
428
429
430
431
432
433

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        noise = torch.randn((1, model.config.in_channels, model.config.resolution, model.config.resolution)).to(
            torch_device
        )
        emb = torch.randn((1, 16, model.config.transformer_dim)).to(torch_device)
        time_step = torch.tensor([10] * noise.shape[0], device=torch_device)

Patrick von Platen's avatar
Patrick von Platen committed
434
        model.to(torch_device)
435
436
437
438
        with torch.no_grad():
            output = model(noise, time_step, emb)

        output, _ = torch.split(output, 3, dim=1)
Patrick von Platen's avatar
Patrick von Platen committed
439
        output_slice = output[0, -1, -3:, -3:].cpu().flatten()
440
        # fmt: off
Patrick von Platen's avatar
Patrick von Platen committed
441
        expected_output_slice = torch.tensor([2.7766, -10.3558, -14.9149, -0.9376, -14.9175, -17.7679, -5.5565, -12.9521, -12.9845])
442
443
444
445
        # fmt: on
        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))


patil-suraj's avatar
patil-suraj committed
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
class UNetLDMModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = UNetLDMModel

    @property
    def dummy_input(self):
        batch_size = 4
        num_channels = 4
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor([10]).to(torch_device)

        return {"x": noise, "timesteps": time_step}

    @property
    def get_input_shape(self):
        return (4, 32, 32)

    @property
    def get_output_shape(self):
        return (4, 32, 32)

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "image_size": 32,
            "in_channels": 4,
            "out_channels": 4,
            "model_channels": 32,
            "num_res_blocks": 2,
            "attention_resolutions": (16,),
            "channel_mult": (1, 2),
            "num_heads": 2,
            "conv_resample": True,
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict
anton-l's avatar
anton-l committed
482

patil-suraj's avatar
patil-suraj committed
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
    def test_from_pretrained_hub(self):
        model, loading_info = UNetLDMModel.from_pretrained("fusing/unet-ldm-dummy", output_loading_info=True)
        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)

        model.to(torch_device)
        image = model(**self.dummy_input)

        assert image is not None, "Make sure output is not None"

    def test_output_pretrained(self):
        model = UNetLDMModel.from_pretrained("fusing/unet-ldm-dummy")
        model.eval()

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        noise = torch.randn(1, model.config.in_channels, model.config.image_size, model.config.image_size)
        time_step = torch.tensor([10] * noise.shape[0])

        with torch.no_grad():
            output = model(noise, time_step)

        output_slice = output[0, -1, -3:, -3:].flatten()
        # fmt: off
        expected_output_slice = torch.tensor([-13.3258, -20.1100, -15.9873, -17.6617, -23.0596, -17.9419, -13.3675, -16.1889, -12.3800])
        # fmt: on

        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))

Patrick von Platen's avatar
Patrick von Platen committed
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
    def test_output_pretrained_spatial_transformer(self):
        model = UNetLDMModel.from_pretrained("fusing/unet-ldm-dummy-spatial")
        model.eval()

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        noise = torch.randn(1, model.config.in_channels, model.config.image_size, model.config.image_size)
        context = torch.ones((1, 16, 64), dtype=torch.float32)
        time_step = torch.tensor([10] * noise.shape[0])

        with torch.no_grad():
            output = model(noise, time_step, context=context)

        output_slice = output[0, -1, -3:, -3:].flatten()
        # fmt: off
        expected_output_slice = torch.tensor([61.3445, 56.9005, 29.4339, 59.5497, 60.7375, 34.1719, 48.1951, 42.6569, 25.0890])
        # fmt: on

        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))

patil-suraj's avatar
patil-suraj committed
536

patil-suraj's avatar
patil-suraj committed
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
class UNetGradTTSModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = UNetGradTTSModel

    @property
    def dummy_input(self):
        batch_size = 4
        num_features = 32
        seq_len = 16

        noise = floats_tensor((batch_size, num_features, seq_len)).to(torch_device)
        condition = floats_tensor((batch_size, num_features, seq_len)).to(torch_device)
        mask = floats_tensor((batch_size, 1, seq_len)).to(torch_device)
        time_step = torch.tensor([10] * batch_size).to(torch_device)

        return {"x": noise, "timesteps": time_step, "mu": condition, "mask": mask}

    @property
    def get_input_shape(self):
        return (4, 32, 16)

    @property
    def get_output_shape(self):
        return (4, 32, 16)

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "dim": 64,
            "groups": 4,
            "dim_mults": (1, 2),
            "n_feats": 32,
            "pe_scale": 1000,
            "n_spks": 1,
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict
anton-l's avatar
anton-l committed
572

patil-suraj's avatar
patil-suraj committed
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
    def test_from_pretrained_hub(self):
        model, loading_info = UNetGradTTSModel.from_pretrained("fusing/unet-grad-tts-dummy", output_loading_info=True)
        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)

        model.to(torch_device)
        image = model(**self.dummy_input)

        assert image is not None, "Make sure output is not None"

    def test_output_pretrained(self):
        model = UNetGradTTSModel.from_pretrained("fusing/unet-grad-tts-dummy")
        model.eval()

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)
anton-l's avatar
anton-l committed
590

patil-suraj's avatar
patil-suraj committed
591
592
593
594
595
596
597
598
599
600
601
602
        num_features = model.config.n_feats
        seq_len = 16
        noise = torch.randn((1, num_features, seq_len))
        condition = torch.randn((1, num_features, seq_len))
        mask = torch.randn((1, 1, seq_len))
        time_step = torch.tensor([10])

        with torch.no_grad():
            output = model(noise, time_step, condition, mask)

        output_slice = output[0, -3:, -3:].flatten()
        # fmt: off
Patrick von Platen's avatar
Patrick von Platen committed
603
        expected_output_slice = torch.tensor([-0.0690, -0.0531, 0.0633, -0.0660, -0.0541, 0.0650, -0.0656, -0.0555, 0.0617])
patil-suraj's avatar
patil-suraj committed
604
605
606
607
608
        # fmt: on

        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))


609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
class NCSNppModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = NCSNpp

    @property
    def dummy_input(self):
        batch_size = 4
        num_channels = 3
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor(batch_size * [10]).to(torch_device)

        return {"x": noise, "timesteps": time_step}

    @property
    def get_input_shape(self):
        return (3, 32, 32)

    @property
    def get_output_shape(self):
        return (3, 32, 32)

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "image_size": 32,
            "ch_mult": [1, 2, 2, 2],
            "nf": 32,
            "fir": True,
            "progressive": "output_skip",
            "progressive_combine": "sum",
            "progressive_input": "input_skip",
            "scale_by_sigma": True,
            "skip_rescale": True,
            "embedding_type": "fourier",
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict

    def test_from_pretrained_hub(self):
        model, loading_info = NCSNpp.from_pretrained("fusing/cifar10-ncsnpp-ve", output_loading_info=True)
        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)

        model.to(torch_device)
        image = model(**self.dummy_input)

        assert image is not None, "Make sure output is not None"

    def test_output_pretrained_ve_small(self):
        model = NCSNpp.from_pretrained("fusing/ncsnpp-cifar10-ve-dummy")
        model.eval()
        model.to(torch_device)

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        batch_size = 4
        num_channels = 3
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor(batch_size * [10]).to(torch_device)

        with torch.no_grad():
            output = model(noise, time_step)

        output_slice = output[0, -3:, -3:, -1].flatten().cpu()
        # fmt: off
        expected_output_slice = torch.tensor([3.1909e-07, -8.5393e-08, 4.8460e-07, -4.5550e-07, -1.3205e-06, -6.3475e-07, 9.7837e-07, 2.9974e-07, 1.2345e-06])
        # fmt: on

        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))

    def test_output_pretrained_ve_large(self):
        model = NCSNpp.from_pretrained("fusing/ncsnpp-ffhq-ve-dummy")
        model.eval()
        model.to(torch_device)

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        batch_size = 4
        num_channels = 3
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor(batch_size * [10]).to(torch_device)

        with torch.no_grad():
            output = model(noise, time_step)

        output_slice = output[0, -3:, -3:, -1].flatten().cpu()
        # fmt: off
        expected_output_slice = torch.tensor([-8.3299e-07, -9.0431e-07, 4.0585e-08, 9.7563e-07, 1.0280e-06, 1.0133e-06, 1.4979e-06, -2.9716e-07, -6.1817e-07])
        # fmt: on

        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))

    def test_output_pretrained_vp(self):
        model = NCSNpp.from_pretrained("fusing/ddpm-cifar10-vp-dummy")
        model.eval()
        model.to(torch_device)

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        batch_size = 4
        num_channels = 3
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor(batch_size * [10]).to(torch_device)

        with torch.no_grad():
            output = model(noise, time_step)

        output_slice = output[0, -3:, -3:, -1].flatten().cpu()
        # fmt: off
        expected_output_slice = torch.tensor([-3.9086e-07, -1.1001e-05, 1.8881e-06, 1.1106e-05, 1.6629e-06, 2.9820e-06, 8.4978e-06, 8.0253e-07, 1.5435e-06])
        # fmt: on

        self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))


736
737
738
739
class PipelineTesterMixin(unittest.TestCase):
    def test_from_pretrained_save_pretrained(self):
        # 1. Load models
        model = UNetModel(ch=32, ch_mult=(1, 2), num_res_blocks=2, attn_resolutions=(16,), resolution=32)
Patrick von Platen's avatar
Patrick von Platen committed
740
        schedular = DDPMScheduler(timesteps=10)
741

Patrick von Platen's avatar
Patrick von Platen committed
742
        ddpm = DDPMPipeline(model, schedular)
743
744
745

        with tempfile.TemporaryDirectory() as tmpdirname:
            ddpm.save_pretrained(tmpdirname)
Patrick von Platen's avatar
Patrick von Platen committed
746
            new_ddpm = DDPMPipeline.from_pretrained(tmpdirname)
Patrick von Platen's avatar
Patrick von Platen committed
747
748

        generator = torch.manual_seed(0)
749

patil-suraj's avatar
patil-suraj committed
750
        image = ddpm(generator=generator)
Patrick von Platen's avatar
Patrick von Platen committed
751
        generator = generator.manual_seed(0)
patil-suraj's avatar
patil-suraj committed
752
        new_image = new_ddpm(generator=generator)
753
754
755
756
757
758
759

        assert (image - new_image).abs().sum() < 1e-5, "Models don't give the same forward pass"

    @slow
    def test_from_pretrained_hub(self):
        model_path = "fusing/ddpm-cifar10"

Patrick von Platen's avatar
Patrick von Platen committed
760
        ddpm = DDPMPipeline.from_pretrained(model_path)
761
762
763
764
765
        ddpm_from_hub = DiffusionPipeline.from_pretrained(model_path)

        ddpm.noise_scheduler.num_timesteps = 10
        ddpm_from_hub.noise_scheduler.num_timesteps = 10

Patrick von Platen's avatar
Patrick von Platen committed
766
        generator = torch.manual_seed(0)
767

patil-suraj's avatar
patil-suraj committed
768
        image = ddpm(generator=generator)
Patrick von Platen's avatar
Patrick von Platen committed
769
        generator = generator.manual_seed(0)
patil-suraj's avatar
patil-suraj committed
770
        new_image = ddpm_from_hub(generator=generator)
771
772

        assert (image - new_image).abs().sum() < 1e-5, "Models don't give the same forward pass"
Patrick von Platen's avatar
Patrick von Platen committed
773
774
775
776
777
778

    @slow
    def test_ddpm_cifar10(self):
        generator = torch.manual_seed(0)
        model_id = "fusing/ddpm-cifar10"

Patrick von Platen's avatar
Patrick von Platen committed
779
        unet = UNetModel.from_pretrained(model_id)
Patrick von Platen's avatar
Patrick von Platen committed
780
        noise_scheduler = DDPMScheduler.from_config(model_id)
Patrick von Platen's avatar
Patrick von Platen committed
781
        noise_scheduler = noise_scheduler.set_format("pt")
Patrick von Platen's avatar
Patrick von Platen committed
782

Patrick von Platen's avatar
Patrick von Platen committed
783
        ddpm = DDPMPipeline(unet=unet, noise_scheduler=noise_scheduler)
Patrick von Platen's avatar
Patrick von Platen committed
784
785
786
787
788
789
790
791
792
793
794
795
796
        image = ddpm(generator=generator)

        image_slice = image[0, -1, -3:, -3:].cpu()

        assert image.shape == (1, 3, 32, 32)
        expected_slice = torch.tensor([0.2250, 0.3375, 0.2360, 0.0930, 0.3440, 0.3156, 0.1937, 0.3585, 0.1761])
        assert (image_slice.flatten() - expected_slice).abs().max() < 1e-2

    @slow
    def test_ddim_cifar10(self):
        generator = torch.manual_seed(0)
        model_id = "fusing/ddpm-cifar10"

Patrick von Platen's avatar
Patrick von Platen committed
797
        unet = UNetModel.from_pretrained(model_id)
Patrick von Platen's avatar
Patrick von Platen committed
798
        noise_scheduler = DDIMScheduler(tensor_format="pt")
Patrick von Platen's avatar
Patrick von Platen committed
799

Patrick von Platen's avatar
Patrick von Platen committed
800
        ddim = DDIMPipeline(unet=unet, noise_scheduler=noise_scheduler)
Patrick von Platen's avatar
Patrick von Platen committed
801
802
803
804
805
        image = ddim(generator=generator, eta=0.0)

        image_slice = image[0, -1, -3:, -3:].cpu()

        assert image.shape == (1, 3, 32, 32)
Patrick von Platen's avatar
Patrick von Platen committed
806
807
808
        expected_slice = torch.tensor(
            [-0.7383, -0.7385, -0.7298, -0.7364, -0.7414, -0.7239, -0.6737, -0.6813, -0.7068]
        )
Patrick von Platen's avatar
Patrick von Platen committed
809
        assert (image_slice.flatten() - expected_slice).abs().max() < 1e-2
patil-suraj's avatar
patil-suraj committed
810

Patrick von Platen's avatar
Patrick von Platen committed
811
812
813
814
815
816
817
818
    @slow
    def test_pndm_cifar10(self):
        generator = torch.manual_seed(0)
        model_id = "fusing/ddpm-cifar10"

        unet = UNetModel.from_pretrained(model_id)
        noise_scheduler = PNDMScheduler(tensor_format="pt")

Patrick von Platen's avatar
Patrick von Platen committed
819
        pndm = PNDMPipeline(unet=unet, noise_scheduler=noise_scheduler)
Patrick von Platen's avatar
Patrick von Platen committed
820
821
822
823
824
825
826
827
828
829
        image = pndm(generator=generator)

        image_slice = image[0, -1, -3:, -3:].cpu()

        assert image.shape == (1, 3, 32, 32)
        expected_slice = torch.tensor(
            [-0.7888, -0.7870, -0.7759, -0.7823, -0.8014, -0.7608, -0.6818, -0.7130, -0.7471]
        )
        assert (image_slice.flatten() - expected_slice).abs().max() < 1e-2

patil-suraj's avatar
patil-suraj committed
830
    @slow
patil-suraj's avatar
patil-suraj committed
831
    @unittest.skip("Skipping for now as it takes too long")
patil-suraj's avatar
patil-suraj committed
832
833
    def test_ldm_text2img(self):
        model_id = "fusing/latent-diffusion-text2im-large"
Patrick von Platen's avatar
Patrick von Platen committed
834
        ldm = LatentDiffusionPipeline.from_pretrained(model_id)
patil-suraj's avatar
patil-suraj committed
835
836
837
838
839
840
841
842
843

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.manual_seed(0)
        image = ldm([prompt], generator=generator, num_inference_steps=20)

        image_slice = image[0, -1, -3:, -3:].cpu()

        assert image.shape == (1, 3, 256, 256)
        expected_slice = torch.tensor([0.7295, 0.7358, 0.7256, 0.7435, 0.7095, 0.6884, 0.7325, 0.6921, 0.6458])
Patrick von Platen's avatar
update  
Patrick von Platen committed
844
        assert (image_slice.flatten() - expected_slice).abs().max() < 1e-2
845

patil-suraj's avatar
patil-suraj committed
846
847
848
849
850
851
852
    @slow
    def test_ldm_text2img_fast(self):
        model_id = "fusing/latent-diffusion-text2im-large"
        ldm = LatentDiffusionPipeline.from_pretrained(model_id)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.manual_seed(0)
853
        image = ldm([prompt], generator=generator, num_inference_steps=1)
patil-suraj's avatar
patil-suraj committed
854
855
856
857

        image_slice = image[0, -1, -3:, -3:].cpu()

        assert image.shape == (1, 3, 256, 256)
patil-suraj's avatar
patil-suraj committed
858
        expected_slice = torch.tensor([0.3163, 0.8670, 0.6465, 0.1865, 0.6291, 0.5139, 0.2824, 0.3723, 0.4344])
patil-suraj's avatar
patil-suraj committed
859
860
        assert (image_slice.flatten() - expected_slice).abs().max() < 1e-2

anton-l's avatar
anton-l committed
861
862
863
    @slow
    def test_glide_text2img(self):
        model_id = "fusing/glide-base"
Patrick von Platen's avatar
Patrick von Platen committed
864
        glide = GlidePipeline.from_pretrained(model_id)
anton-l's avatar
anton-l committed
865
866
867
868
869
870
871
872
873
874
875

        prompt = "a pencil sketch of a corgi"
        generator = torch.manual_seed(0)
        image = glide(prompt, generator=generator, num_inference_steps_upscale=20)

        image_slice = image[0, :3, :3, -1].cpu()

        assert image.shape == (1, 256, 256, 3)
        expected_slice = torch.tensor([0.7119, 0.7073, 0.6460, 0.7780, 0.7423, 0.6926, 0.7378, 0.7189, 0.7784])
        assert (image_slice.flatten() - expected_slice).abs().max() < 1e-2

Patrick von Platen's avatar
Patrick von Platen committed
876
877
878
    @slow
    def test_grad_tts(self):
        model_id = "fusing/grad-tts-libri-tts"
Patrick von Platen's avatar
Patrick von Platen committed
879
        grad_tts = GradTTSPipeline.from_pretrained(model_id)
880
881
        noise_scheduler = GradTTSScheduler()
        grad_tts.noise_scheduler = noise_scheduler
Patrick von Platen's avatar
Patrick von Platen committed
882
883

        text = "Hello world, I missed you so much."
Patrick von Platen's avatar
Patrick von Platen committed
884
        generator = torch.manual_seed(0)
Patrick von Platen's avatar
Patrick von Platen committed
885
886

        # generate mel spectograms using text
Patrick von Platen's avatar
Patrick von Platen committed
887
        mel_spec = grad_tts(text, generator=generator)
Patrick von Platen's avatar
Patrick von Platen committed
888

Patrick von Platen's avatar
Patrick von Platen committed
889
890
        assert mel_spec.shape == (1, 80, 143)
        expected_slice = torch.tensor(
Patrick von Platen's avatar
Patrick von Platen committed
891
            [-6.7584, -6.8347, -6.3293, -6.6437, -6.7233, -6.4684, -6.1187, -6.3172, -6.6890]
Patrick von Platen's avatar
Patrick von Platen committed
892
        )
Patrick von Platen's avatar
Patrick von Platen committed
893
        assert (mel_spec[0, :3, :3].cpu().flatten() - expected_slice).abs().max() < 1e-2
Patrick von Platen's avatar
Patrick von Platen committed
894

Patrick von Platen's avatar
Patrick von Platen committed
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
    @slow
    def test_score_sde_ve_pipeline(self):
        torch.manual_seed(0)

        model = NCSNpp.from_pretrained("fusing/ffhq_ncsnpp")
        scheduler = ScoreSdeVeScheduler.from_config("fusing/ffhq_ncsnpp")

        sde_ve = ScoreSdeVePipeline(model=model, scheduler=scheduler)

        image = sde_ve(num_inference_steps=2)

        expected_image_sum = 3382810112.0
        expected_image_mean = 1075.366455078125

        assert (image.abs().sum() - expected_image_sum).abs().cpu().item() < 1e-2
        assert (image.abs().mean() - expected_image_mean).abs().cpu().item() < 1e-4

Patrick von Platen's avatar
Patrick von Platen committed
912
913
914
    @slow
    def test_score_sde_vp_pipeline(self):

Patrick von Platen's avatar
Patrick von Platen committed
915
916
        model = NCSNpp.from_pretrained("fusing/cifar10-ddpmpp-vp")
        scheduler = ScoreSdeVpScheduler.from_config("fusing/cifar10-ddpmpp-vp")
Patrick von Platen's avatar
Patrick von Platen committed
917
918
919
920
921
922
923
924
925
926
927
928

        sde_vp = ScoreSdeVpPipeline(model=model, scheduler=scheduler)

        torch.manual_seed(0)
        image = sde_vp(num_inference_steps=10)

        expected_image_sum = 4183.2012
        expected_image_mean = 1.3617

        assert (image.abs().sum() - expected_image_sum).abs().cpu().item() < 1e-2
        assert (image.abs().mean() - expected_image_mean).abs().cpu().item() < 1e-4

929
930
931
932
    def test_module_from_pipeline(self):
        model = DiffWave(num_res_layers=4)
        noise_scheduler = DDPMScheduler(timesteps=12)

Patrick von Platen's avatar
Patrick von Platen committed
933
        bddm = BDDMPipeline(model, noise_scheduler)
934
935
936
937
938
939
940

        # check if the library name for the diffwave moduel is set to pipeline module
        self.assertTrue(bddm.config["diffwave"][0] == "pipeline_bddm")

        # check if we can save and load the pipeline
        with tempfile.TemporaryDirectory() as tmpdirname:
            bddm.save_pretrained(tmpdirname)
Patrick von Platen's avatar
Patrick von Platen committed
941
            _ = BDDMPipeline.from_pretrained(tmpdirname)
942
            # check if the same works using the DifusionPipeline class
943
            _ = DiffusionPipeline.from_pretrained(tmpdirname)